Add Tracking Cloth Deformation evaluation on gpuserver6000

[FlorianPfaff]

User-requested public-data evaluation

Use the already installed Zenodo 14644526 dataset on runner label gpuserver6000 (workstation2), defaulting to /home/github-runner/.cache/datasets/tracking-cloth-deformation-v1-zenodo-14644526.

This is executable evaluation code, not another acquisition protocol: a small NumPy spring-mesh physical baseline, deterministic residual/state-injection/MAP controls, source-only generalized Bayesian parameter averaging, and a source-frozen empirical accept/fallback rule.

Workflow

One maintained entry point with inventory, source_only (default), and evaluate modes. Pull requests run synthetic tests on GitHub-hosted runners only. Real-data execution is restricted to an explicit workflow_dispatch of main; the self-hosted gpuserver6000 job is skipped on PRs. Uses read-only repository permissions, pinned actions, isolated venv/scratch, and no raw-recording or trajectory-array upload.

• Verify the published archive MD5, ZIP integrity, and exact extracted CSV bytes.
• Fit only the complete 32-recording shaking source factorial.
• Freeze all 32 twisting predictions; upload the complete seal before a separate scoring step.
• Condition on recorded future driven-corner trajectories, not unlogged robot commands.
• Score only free-marker futures after a 1-second prefix, over a fixed 5-second horizon.
• Report all seven arms, per-record/per-specimen point errors, coordinate NLL, coverage with width, harmful accepted updates, exact fallback, and paired specimen/material bootstrap summaries.

The remaining 56 collision recordings are not numerically accessed. No case substitution, target-side fitting, or partial-roster pooled result.

Evidence and modeling boundary

Classification: operational prerequisite and scored public-data diagnostic for the external shake-to-twist question. This is explicitly a pilot, not reproduced PhysTwin/clothilde-sim/FEM, calibrated material identification, a safety certificate, or fresh confirmation. Source guards use four-fold leave-one-speed/grasp-recording-out results. Initial marker-grid assumptions must pass qualification. The eight specimen/four material counts are kept separate from frames and coordinates.

Zenodo CC BY-SA metadata conflicts with included CC BY-NC-SA; the included noncommercial policy and license text are retained pending author clarification. Cache is never modified. claims.json, prior evidence and the paper are unchanged.

Validation

The branch-side repair run passed 18 synthetic tests before push. On the final PR head, Tracking-cloth synthetic contracts passes (tests, Ruff, and format), and Changed source and workflow preflight passes (including workflow lifecycle policy). Repository-wide Tests and Security scanning run independently. No real-data execution/result is claimed yet; the gpuserver6000 PR job is intentionally skipped.

The existing Cloth Sim2Real workflow binds another dataset and historical protocol, so it is deliberately not repurposed.